Former Intel CEO on What Went Wrong, What's Next + Lovable CEO on the Real Promise of Vibe Coding

Source: Authorized podcast transcript supplied by the user. The transcript contains two main interview segments separated by sponsor reads.

Speakers identifiable from the transcript:

• Pat — identified by the host as “Pat”; the discussion and biographical details identify him as the former longtime Intel executive being interviewed. Exact full-name presentation should be checked before publication.
• Jason — the host/interviewer is explicitly addressed as “Jason” near the end of the second interview.
• Anton / “Oseka” — founder of Lovable. The transcript addresses him as Anton but renders the founder’s name once as “Oseka.” Exact spelling should be verified.
• Other voices consist mainly of short host/interviewer interjections.

Source context: The first interview focuses on Intel’s rise and decline, Apple Silicon, Nvidia, TSMC, semiconductor manufacturing, Taiwan risk, AI infrastructure, valuations and quantum computing. The second focuses on Lovable, AI-generated software, bespoke enterprise applications, security, AI model strategy, software economics and the changing role of human judgment.

Transcript-quality issues: The transcript is generally understandable but contains numerous probable transcription errors or uncertain proper nouns, including dates, semiconductor terminology, AI-company names, quantum terminology, company names in the Lovable examples, and the name of an Anthropic model/product. These have not been silently corrected below. They are identified in Section D.

The transcript also includes promotional reads for Airwallex and Plaud. Those are separated from the substantive interviews below rather than treated as interview evidence.

The host opens by noting that Pat spent 34 years at Intel and describes Intel as historically one of the greatest American companies, before being overtaken or severely challenged by Nvidia, TSMC and, to some degree, Apple.

The host recalls the cultural power of the “Intel Inside” era, when consumers paid attention to the processor brand—particularly Pentium—when buying computers.

The interview is framed as a post-mortem: Intel achieved enormous success, later made critical mistakes, and now appears to be recovering to some extent. The host asks what went wrong, what went right, and what strategic decisions should have been different.

Pat says Intel was personally formative for him because he joined at 18 years old and effectively grew up inside the company.

He describes Andy Grove, Noyes, Barrett and other early Intel leaders as mentors and people he deeply admired.

A central feature of the Intel he remembers was its technical depth. He says that when he first joined the executive staff, roughly 15 of the 20 people in the room were PhDs.

Pat argues that one of Intel’s fundamental failures occurred when leadership shifted from technical people toward business-oriented executives.

His reasoning is organizational as well as technological:

• A business-oriented CEO tends to promote other business-oriented executives.
• Those executives then reproduce the same type of leadership further down the organization.
• Over time, technical judgment becomes less central to strategic decisions.
• In a company whose fundamental business is technology, that shift ultimately affects investment choices, staffing and the technologies the company is willing to pursue.

Pat states that when he became CEO in “2001”, he was the first technical leader in roughly 15 years. That date appears questionable from the transcript and requires verification rather than silent correction.

He contrasts this with successful modern technology companies. He says they are often founder-led, but argues that founder status itself is not essential. He cites Satya [transcript wording: “Sacha”] and Sundar as examples of non-founders who nevertheless have deep technical backgrounds.

His larger argument is that when executives are making technical decisions involving billions of dollars, a spreadsheet alone cannot tell them whether an investment is right. Something can look financially unattractive in isolation but still be essential because of the underlying direction of technology.

Pat: “Fundamentally, this is a technology business.”

He therefore believes technology companies need technologists at the top, who hire technical leaders beneath them, who in turn recruit the strongest technical people available.

Pat says Intel returned approximately $100 billion to shareholders through dividends and stock buybacks during the five or six years before he returned.

He says he would have greatly preferred to have had that additional $100 billion available on Intel’s balance sheet.

The host suggests one possible missed opportunity was making chips for the iPhone.

Pat instead emphasizes the manufacturing-investment problem. He says Intel had not built a new factory in a decade when he arrived/returned.

He questions how a semiconductor company could fail to keep building factories or acquiring EUV machines.

His point is that these investments might have looked unattractive to a financial manager when viewed individually, but a technologist would understand why the direction of semiconductor technology required them.

Pat acknowledges that all leaders and businesses make both good and bad decisions. His argument is not that every mistake can be eliminated, but that Intel’s deeper mistake was losing sight of the fact that it was fundamentally a technology company.

The host draws a parallel with Apple, arguing that Apple has also spent enormous amounts on stock buybacks and dividends over approximately 15 years.

He describes Apple as an enormous holder of capital and argues that it has generally made relatively small acquisitions rather than transformative purchases.

He cites Beats as what he believes was one of Apple’s largest acquisitions and says Apple used it partly to gain access to demographic segments where it lacked reach.

The host characterizes this conservative acquisition strategy as a missed opportunity and argues that Apple could have used its capital for more ambitious investments.

He then asks Pat about Steve Jobs’ decision around 2008–2009 to begin developing Apple’s own silicon and whether Intel knew that this effort was underway.

Pat describes Steve Jobs as both an “incredible leader” and a “ruthless leader.”

He says Jobs could be very difficult and refers listeners to Walter Isaacson’s biography.

Pat says he had many conversations with Jobs over the years.

When Apple moved to Intel and the “Centrino chip” [source wording; verify technical description], Pat says it was a major event.

Apple was a demanding customer. Pat says it continually pushed Intel to:

• make chips smaller;
• reduce their power consumption;
• and continue improving the technology.

According to Pat, Jobs began developing an internal alternative when he stopped being confident that Intel would continue meeting Apple’s needs.

Apple first acquired small companies—Pat mentions something transcribed as “P semi”—and gradually built semiconductor competence internally.

His interpretation is that Apple’s semiconductor program did not begin as one giant bet. Instead, Jobs repeatedly started small internal programs that created capabilities the company might need later.

Pat offers an earlier example involving Apple’s operating system.

When Intel discussed helping Apple port its operating system from Power architecture to Intel’s x86 platform, Intel believed its own compiler and operating-system expertise would be valuable.

Jobs reportedly responded that Apple had already been porting the operating system to x86 for the previous four releases.

Pat says he was shocked. Jobs had quietly prepared the underlying technology for years before Apple needed to make the platform transition.

Pat sees Apple Silicon as the same strategic pattern: build competence before the strategic need becomes unavoidable.

Jobs’ logic, in Pat’s interpretation, was not simply that Intel had failed as a supplier. It was that Apple eventually concluded it could serve itself better.

By controlling both the operating system and semiconductor design, Apple could optimize the whole system instead of relying on silicon primarily optimized for another ecosystem such as Windows.

Pat’s characterization of Apple’s decision: “It was never that kind of thing that said, you know, you failed as a supplier. No, I can supply myself better.”

The host next asks how much of Nvidia’s success came from skill versus luck, noting that GPUs unexpectedly became valuable for cryptocurrency and later AI.

Pat says Jensen initially focused on building high-throughput computing machines.

During Intel’s period of dominance in CPUs, Intel people tended to dismiss Nvidia’s products as graphics devices useful mainly to gamers.

The conceptual hierarchy at the time was, in Pat’s description, the important big CPU versus the smaller, less important GPU.

The shift began when Nvidia built a serious software stack around its hardware.

The transcript refers to “Kuta” and “SimT” technology; these terms require verification.

Generation after generation, Nvidia kept making that software and platform more capable.

Pat compares this incremental process with Jobs’ approach: improve something continuously until its possible applications expand.

A major turning point came when Japanese high-performance-computing researchers began experimenting with graphics cards for HPC workloads.

That demonstrated that GPUs were no longer merely graphics processors. They were computationally dense devices capable of addressing other demanding workloads.

Pat believes this was a defining realization for Nvidia: the product could become a general-purpose computing platform.

AI itself had already experienced repeated periods of disappointment. Pat describes the field as having gone through perhaps its “fifth nuclear winter.”

Even when AI seemed unlikely to succeed commercially, however:

• the research community continued working;
• Nvidia’s software continued improving;
• and the hardware/software platform became increasingly useful.

Pat notes that Intel had its own project, Larrabee, intended to take x86 architecture in a broadly similar direction.

He says Larrabee was killed approximately one week after his first departure from Intel.

He believes the technological world could have developed very differently had that project continued.

The host interprets Nvidia as an example of the compounding effects of continuous innovation, basic research and technological experimentation.

He invokes William Gibson’s idea that “the street finds its own use for technology.”

The host’s reasoning is that Nvidia did not originally build GPUs specifically for Bitcoin or the subsequent AI boom. Instead, once the technology existed and was powerful enough, outside communities discovered new applications for it.

The host identifies TSMC as another major strategic challenge Intel failed to anticipate.

Pat says TSMC began with a remarkable vision: become “the factory for the industry.”

He emphasizes the scale required. Advanced semiconductor factories can cost approximately $20 billion to $30 billion, while also requiring continual engineering and capital investment.

Intel historically operated as an IDM — integrated design and manufacturing company.

Intel manufactured its own CPUs and reused its manufacturing capabilities for chipsets and other internal products.

It did not, however, build its factories and processes as standardized services designed for a large range of third-party chip designers.

Pat contrasts this with the ecosystem required by a foundry:

• standardized processes;
• PDKs;
• broadly usable design tools;
• interfaces that allow many external chip designers to use the same manufacturing platform.

Intel’s systems were highly proprietary.

Pat also says he worked on early foundations of EDA inside Intel, mentioning early place-and-route systems, standard cells and high-level description language work.

TSMC’s conceptual breakthrough was to separate chip design from manufacturing.

Its position was essentially: it did not matter who designed the chip or what the chip did; TSMC would be the manufacturing partner.

Initially, the market for this model was so small relative to Intel’s own CPU business that Intel did not consider it important.

But TSMC made steady progress over a long period.

Pat specifically credits Apple as an important customer that helped push TSMC into becoming a highly significant advanced manufacturer.

Pat says that when he returned to Intel in “2001” [date as transcribed; verify], TSMC was producing approximately five times as many wafers as Intel.

The importance of that number, in his framing, is that this was not a marginal lead of 10% or 20%; the foundry model had grown to several multiples of Intel’s manufacturing scale.

The semiconductor industry eventually organized itself broadly around the foundry model, with Intel and memory manufacturers described as the main exceptions.

Pat says that when he returned to Intel, one of his core strategic conclusions was that Intel itself had to become a foundry.

The host says the wafer ratio may now be closer to seven to one, TSMC versus Intel.

The host asks whether the United States can realistically reproduce enough semiconductor capacity domestically to reduce dependence on Taiwan.

The question is framed against the risk of a Chinese blockade or attack and includes a hypothetical scenario in which Taiwanese fabs could be destroyed or engineers relocated.

Pat divides his response into several points.

First, he argues that the CHIPS Act is producing real benefits.

He says that when the effort began, the United States was producing approximately 12% of leading-edge semiconductor capacity and that the figure is now closer to 18%.

He stresses that 18% is still far from 50%, but regards the change as meaningful progress.

He also says:

• Intel is beginning to become a genuine foundry;
• TSMC's U.S. factories are operating at scale;
• Samsung is also part of the U.S. manufacturing picture.

He characterizes Intel and TSMC’s progress as particularly meaningful.

The source again contains a questionable reference to “2001” in connection with Pat returning and the CHIPS Act chronology. This must be verified before publication.

Pat then gives what he calls the “ugly” side of the semiconductor-supply-chain issue.

He claims Taiwan has less than three weeks of energy reserves, citing a Wall Street Journal article that he says appeared roughly two weeks before the interview.

According to Pat:

• a blockade that stopped incoming energy supplies could cause the island to brown out after approximately three weeks;
• Taiwan depends on imported oil and LNG;
• semiconductor fabs cannot simply restart immediately once power is restored;
• if a fab is shut down, he says it can take 90 days to come fully back online.

He claims the global economic impact of Taiwan losing power could be greater than the Great Depression.

His key strategic point is that a catastrophic semiconductor interruption would not necessarily require an invasion or even a shot being fired. An energy blockade by itself could have enormous consequences.

He says this possibility should make policymakers and business leaders take semiconductor resilience much more seriously.

Pat adds an important qualification: he does not claim to have access to all classified intelligence or to be sitting in a national-security situation room.

Nevertheless, he claims China has blockaded or conducted blockade-like activity in the Taiwan Strait seven times over the previous four years.

The host agrees that Chinese exercises have been provocative.

Pat does not claim certainty about whether a crisis will occur in 2027, 2030, 2035 or another year. His argument is instead that Chinese intentions and pressure have been sufficiently persistent that supply-chain resilience cannot be treated as optional.

He says he continues devoting substantial time and energy to this problem.

His conclusion is that progress is being made, but the United States and its partners need to move faster and more meaningfully.

The host compares the current AI buildout with previous infrastructure waves such as:

• personal computers;
• servers;
• the internet.

He calls the AI expansion the “buildout to end all buildouts” because of the enormous expected demand for data centers, chips and inference.

He asks whether the sector is in a bubble and identifies two possible risks:

1. too much infrastructure could be built;
2. AI might fail to solve enough valuable problems to justify the enormous token and infrastructure supply.

The host notes that extraordinarily high company valuations create pressure if spending continues to exceed revenue.

Pat agrees that valuations have become extraordinary and acknowledges that markets can lose enthusiasm for companies that fail to earn enough relative to what they spend.

Pat argues that AI has an unusual constraint that could prevent infrastructure speculation from becoming completely detached from reality: energy supply.

He says global energy capacity is expanding at roughly 4%–5%.

By contrast, he describes U.S. energy-capacity growth during a prior decade as only about 1%, which he characterizes very negatively.

Energy infrastructure is now being expanded.

His key point is that a company will not purchase enormous numbers of GPUs or build data centers unless power is available to operate them.

That creates a physical upper bound on the speed at which AI infrastructure can be built.

Pat says this gives him some comfort because speculative investment cannot scale infinitely independent of the electricity system.

Pat distinguishes limits on infrastructure construction from limits on the potential value of AI.

He asks what the incremental value of another token would be if tokens effectively represent additional intelligence.

His answer is that the economic value could be extremely large, because more intelligence could improve:

• supply-chain management;
• finance;
• logistics;
• operational efficiency;
• numerous other economic activities.

He also points to labor shortages in developed countries as another reason additional machine intelligence could be valuable.

For these reasons, Pat describes himself as an optimist.

He predicts that AI infrastructure is not a two- or three-year buildout but a multi-decade buildout, specifically saying “a couple of decades.”

He says one major technical objective should be making AI approximately 10,000 times better.

He also calls for reducing the cost and energy requirements per token by approximately five orders of magnitude.

His goal is to reach a point where sharply lower costs cause AI consumption to explode through what the transcript calls “Jevin’s law.”

The host similarly refers to “Jevin’s paradox.” The terminology should be verified.

The host says users have already experienced a version of this: as AI tokens become cheaper and tools become better, people use them dramatically more, until they eventually confront the total bill and begin asking about ROI.

The speakers mention several inference-chip companies, with the transcript rendering names including “cerebrous,” “rock,” and “Dematrix.” These names should be checked.

Pat argues that enormous improvements in token economics, availability and energy efficiency would support an exceptional period of technological development over the coming decades.

Pat makes an unusually expansive claim about the period ahead.

He says there has never been a better time in human history to be a technologist.

He predicts progress in areas including:

• chemistry;
• language;
• new materials;
• new forms of human/computer interaction;
• cancer;
• poverty reduction.

He views technologists as occupying a uniquely influential position in creating these advances.

Pat: “There is not a better time to be alive than the one that we're in right now.”

This is explicitly Pat’s optimistic prediction and worldview, not an independently established conclusion.

The host returns to the issue of valuations.

He notes that the dot-com boom eventually produced a major disconnect between market valuations and underlying business performance.

He argues that at least some current AI companies differ because they have substantial real revenue.

He mentions ElevenLabs at approximately $600 million in revenue and Lovable at roughly $500 million–$600 million, according to his figures.

Pat agrees that current AI businesses often have genuine revenue and real margins.

However, he warns that excessively high multiples still create the conditions for market corrections.

Pat sees periodic corrections as healthy because they keep earnings multiples and expectations from becoming completely detached from underlying economics.

His long-term optimism does not imply a smooth market trajectory.

He predicts roughly two decades of technological progress but expects numerous disruptions and sharp corrections during that period.

He says investors should, in effect, be grateful when corrections stop speculative excess from getting too far ahead of itself.

He cites the “SaaS apocalypse” as one example and predicts there will be other “apocalypses” as AI capabilities disrupt existing industries.

Pat says the technological story becomes even more significant when three computing paradigms converge:

1. classical computing;
2. AI computing;
3. quantum computing.

He calls this combination the “Trinity of Computing.”

In his view, the combination will enable capabilities that no single computing model can deliver on its own.

The host jokes that quantum computing has seemed “five years away” for roughly 25 years and asks when it will finally become useful.

Pat answers: “This decade.”

When the host asks whether it will therefore be meaningful by 2030, Pat says yes.

He predicts quantum systems will begin solving problems that cannot practically be computed today.

He highlights:

• chemistry;
• biology;
• logistics and optimization.

As a relatively simple example, he cites problems resembling the traveling salesman problem, where quantum systems could potentially calculate better ways to move something from one location to another.

The transcript then contains an unclear date passage—“20, 2032, 23”—in connection with solving encryption-related problems.

Pat associates that later stage with “Q-Day” implications, meaning a point at which quantum systems could affect existing cryptographic systems.

Despite uncertainty around the exact timing of encryption impacts, he predicts quantum-supremacy-type results across multiple industries during this decade.

Pat argues that several fundamental quantum-computing questions have already been answered.

In the transcript’s terminology, he says:

• “we know how to build cubits”;
• “we know how to error-correct cubits”;
• quantum algorithms now exist;
• the remaining problem is increasingly one of engineering and scale.

The term “cubits” appears likely to be a transcription issue and should be verified.

Asked who will win, Pat identifies himself as a supporter of a company transcribed as “Cy Quantum”, which he says is one of the portfolio companies he is associated with.

He nevertheless says multiple technical approaches are showing credible results.

He names approximately four, five or six modalities, including:

• trapped ions;
• photonic approaches;
• spin approaches.

His conclusion is that the industry no longer appears to face a fundamental problem of finding any viable quantum modality.

He also says error correction has been demonstrated across multiple approaches.

The competitive race is therefore moving toward engineering execution and scaling.

His explicit prediction is meaningful quantum results before 2030.

The host notes that, from the interview date, that is only about 40 months away.

After a Plaud advertisement, the host introduces the founder of Lovable, rendered in the transcript as “Oseka” and later addressed as Anton.

The host says he likes the founder because Lovable has become highly compelling to users, including people working for the host.

He describes Lovable’s mission as enabling almost anyone to build great software.

Anton refines the mission as “empowering humans.”

He identifies two gaps Lovable wants to close:

1. enabling a person to build a product;
2. enabling that person to build a functioning business around the product.

Anton says Lovable has made substantial progress on the first problem and is investing heavily in the second.

He reports:

• approximately 1 million new projects created every week;
• more than 700 million visits per month to applications built through the platform;
• more than 50 million apps created on the platform;
• roughly 20 months in market at the time of the interview.

He says the platform continues seeing very rapid growth in the surface area represented by all those applications.

Anton says Lovable serves not only first-time founders but also enterprise leaders.

He argues that the platform enables enterprise leaders and their teams to move considerably faster.

He attributes some of this success to Lovable being opinionated about:

• how software should be created;
• how software should be operated;
• how applications within a company should connect with one another over time.

Anton says enterprise is currently the area where Lovable is growing fastest.

The host compares Lovable with the no-code/low-code wave from roughly ten years earlier.

He recalls founders in his incubator—often MBAs or non-developers—using early no-code tools.

Those tools demonstrated the promise of non-programmers building software, but the output often:

• looked poor;
• did not work reliably;
• performed slowly.

The host argues that large language models finally supplied enough intelligence to fulfill that earlier promise.

He asks whether Lovable’s real audience is developers or the roughly 95% of society that does not traditionally write software.

Anton says roughly 20% of Lovable users are technical or have some type of engineering background.

He says technical users appreciate the fact that Lovable is opinionated and incorporates architectural best practices directly into the platform.

He gives security and payments as examples.

Lovable attempts to make payment setup secure and can perform security scans after every change, as well as monitoring projects in the background.

Technical users also value Lovable because it acts as a bridge allowing non-technical colleagues to participate directly in product creation.

Anton says approximately four out of five users are non-technical.

These users often begin by creating something in order to discover what the correct product should be, an area where Anton believes Lovable has always been particularly strong.

Anton says the platform has moved beyond prototyping.

Some businesses running on Lovable now generate more than $1 million in revenue, according to him.

He argues that Lovable is therefore being built for the entire spectrum from non-technical individuals to professional engineers and larger companies.

He describes a pattern of “cross-pollination”:

• somebody first discovers Lovable through coworkers at a large company;
• that person uses the platform for a side project;
• the side project begins generating substantial income;
• some side businesses make hundreds of thousands of dollars;
• the employee then becomes a founder.

He treats this as evidence that inexpensive software creation changes not only how companies build tools but who can become an entrepreneur.

The host says the perception of vibe coding changed dramatically within roughly a year.

Previously, people might have described it as useful for:

• mock-ups;
• wireframes;
• early product thinking;
• prototypes.

The host argues that this distinction is disappearing because users can now move directly from an idea to a functional product in one or two days.

The harder problem increasingly begins after initial product creation.

He highlights requirements such as:

• payments;
• security;
• preventing data loss;
• preventing data leakage;
• deployment.

He says these capabilities have improved dramatically in approximately the preceding 12 months.

Anton says many engineers using these systems no longer inspect the code or manually write much of it.

His conclusion is that if professional engineers can create software without directly writing code, traditional engineering credentials are no longer a prerequisite for software creation.

Lovable’s role is to give the resulting project a structured architecture and prevent users from making catastrophic mistakes.

Anton says the platform takes care of or guides users through tasks including:

• payments;
• email;
• discoverability by AI chat engines;
• Google search discoverability;
• data security;
• secure connections to external services.

The aim is that users do not need to understand every underlying implementation detail.

Anton attributes Lovable’s success partly to being strongly opinionated about architecture and security from the beginning while designing for what he calls the “99%.”

The host describes giving his team permission to experiment with a variety of AI software-development tools.

Different employees had previously tried:

• Lovable;
• Cursor;
• Claude Code;
• something transcribed as “co-work”;
• Perplexity Computer.

Some attempts produced interesting results but did not become fully finished products.

The host then describes a project for Founder University, his pre-accelerator.

The team wanted an internal intranet.

He says that ten years earlier he would never have approved the project because he believes it would have cost approximately $500,000.

He would rather have spent that money directly on founders and program growth.

Instead, the person running Founder University built the intranet herself using Lovable in approximately four to eight hours.

She had not received explicit permission to build it.

When the host asked how it was created, she told him she had used Lovable and simply put the subscription on her corporate card.

The result was significant enough that the software then began helping operate the program.

The host says Founder University operates in countries including Saudi Arabia and Japan, where demonstrating economic impact matters.

He asked the employee whether she could build a tool showing the economic impact of the roughly 50 companies in the program.

Using Lovable and additional human prompting, she created a system that considered factors including:

• number of employees at each company;
• taxes paid;
• housing/rent;
• average salary.

The host says this was the kind of software he previously would have been unable or unwilling to afford.

The host believes Lovable costs around $50 per month.

Anton says the business plan is approximately that amount and that pricing starts at $25.

The host estimates that software which might previously have cost him about $500,000 was produced by an employee in roughly four hours.

Even after including employee time and software costs, he estimates the solution can be thought of as costing under $2,000 over a year.

He considers the economic difference extraordinary.

When Anton asks what still needs improvement in the intranet, the host says his central concern is now security and data leakage.

He says his team spoke with Lovable’s people, reviewed the project and came away believing the implementation was secure.

Anton says Lovable is also asking penetration-testing specialists to compare AI development tools and examine security capabilities that users do not directly see.

He emphasizes that the company spends heavily on security and trust.

According to Anton, even free users receive significant background security scanning.

The host summarizes the speed of change by saying that roughly a year earlier AI-generated software was mainly associated with mock-ups; now it has reached functionality, security and viable deployment.

Asked where Lovable will be in another year, Anton says the company has identified another gap beyond creating the product.

Lovable now has a separate product line covering hosting, including both AI-related hosting and conventional software hosting.

Anton claims this hosting business is growing faster than the original software-building product.

The host characterizes it as an AWS competitor.

Anton says Lovable can run a customer’s software while itself working with companies such as AWS underneath the platform.

Anton says users increasingly want Lovable not simply as a development tool but as an AI co-founder.

If a company’s applications and operational tools are hosted on the platform, Lovable can potentially access the business data required to understand how the company is performing.

Anton says Lovable is working with some customers in pre-release on an AI partner that continues working while the human founder is asleep.

The proposed system could return in the morning with:

• possible strategic directions;
• optimizations;
• ways to grow faster;
• ways to serve customers better;
• ways to operate faster or more effectively.

Anton describes this as the evolution from merely generating software to providing operation and intelligence aimed at business outcomes.

The host summarizes the proposition as:

“You come to build the software, but you stay to build the business.”

Anton modifies that slightly to “operate your business.”

Anton says Lovable has long attempted to compound everything it learns.

Whenever the platform makes a mistake, the mistake enters a large engineering system designed to improve future performance.

He sees the resulting compounding intelligence as useful not only for software generation but eventually for customers using Lovable to operate their companies.

The implication is that each user interaction can contribute signals that improve the architecture, agent behavior and business assistance available to future users.

The host asks whether software will eventually become almost completely bespoke.

He gives Slack as an example.

His organization spends roughly $10,000 per year on Slack at a high service level, which he considers good value.

Nevertheless, he has started wondering whether it might now make sense to vibe-code an internal Slack alternative that could integrate more deeply with his company’s own systems.

He extends the question to standard enterprise software such as:

• Salesforce;
• HubSpot;
• Slack;
• Google Workspace/Suite;
• Microsoft Office.

His question is whether inexpensive AI-generated applications will allow companies to replace standardized SaaS with software built specifically around their own processes.

Anton answers with a customer story.

The person’s name is transcribed as “Nanad” and the company as “Narsa”; both should be verified.

The person worked at a large U.S. company and initially used Lovable to build a new product line transcribed as “nurse study,” related to educating nurses.

He built tools including:

• administrative systems;
• scheduling;
• licensing workflows;
• certification management.

Because the company already had access to nurses seeking certification, it was able to turn the system into a marketable product.

The same person then brought Lovable into the organization’s internal back office.

Anton says the company has replaced more than ten existing software tools with bespoke applications.

He claims this saves the organization more than $1 million per year.

Anton does not predict that every standardized SaaS product disappears.

He says bespoke software makes sense for at least two reasons:

1. cost savings, as in the example above;
2. special requirements that existing general-purpose products cannot meet precisely.

At the same time, he expects Lovable applications to continue interoperating with standard enterprise tools.

He says Lovable can guide users through connections with:

• Google’s suite of products;
• Microsoft products, with broader support coming soon;
• Slack.

Anton stresses the importance of showing users exactly how data flows between systems so they can avoid granting the wrong people access to sensitive information.

He says companies may therefore continue using systems such as Salesforce and HubSpot underneath their infrastructure while placing a bespoke interface created with Lovable on top.

The host observes that new frontier AI models are simultaneously potential competitors to Lovable and technologies Lovable can use internally.

He notes that critics have repeatedly predicted Lovable’s death whenever a stronger foundation model is released.

He also references extraordinary revenue growth, saying Lovable had moved through successive hundred-million-dollar levels.

The host estimates the business at approximately $400 million in revenue.

Anton replies:

“We reached 500 in May.”

The context suggests a financial run rate or revenue figure, but the exact metric and units are not restated and should be verified.

Anton explains that Lovable’s model strategy is to use whatever produces the best results for customers rather than commit to a single AI provider.

Anton says Lovable uses multiple AI models.

A user request is routed to whichever model appears most suitable for that specific task.

Those models include:

• commercial frontier models from several vendors;
• increasingly, open-weight models.

When a request is sent to Lovable’s own model, the company can improve that model specifically for its agent architecture.

Anton says the areas of improvement include:

• mistakes the model makes;
• deciding which tool to call;
• deciding which integration to create;
• guiding users toward successful business outcomes.

The host interprets this as Lovable being strongly committed to open-source/open-model technology.

Anton qualifies that interpretation.

Lovable maintains multiple partnerships, including relationships with large AI labs, while also using its own models where they can deliver faster performance and lower costs without reducing customer quality.

Anton says Lovable has a strong research team in Stockholm working on post-training.

He says the company is applying current best practices and significantly expanding the team.

He also frames this as strategically important for the European AI ecosystem, because he believes Europe should retain this kind of technical capability locally.

The host asks whether Lovable uses outside data-labeling or data-training companies to identify common business patterns and build proprietary training data.

Anton explains that Lovable focuses directly on the mistakes its models currently make.

The company:

1. identifies model mistakes;
2. ranks them according to customer impact;
3. creates data sets around the most important problems;
4. applies reinforcement-learning techniques specifically where frontier models are still failing.

Anton presents this as a targeted feedback system driven by real customer outcomes rather than simply collecting generic data.

Anton says Lovable receives an enormous amount of token activity because approximately one million new projects are built each week.

The host notes that the company must therefore be consuming an enormous number of AI tokens.

Anton agrees.

He says this usage generates valuable signals for improving both:

• the overall agent harness;
• internal skills developed over approximately two years.

Anton describes those skills as knowledge the agent can call on, including lessons from software engineers about how high-quality software should be built.

Both the agent system and its skills are modified every week based on accumulated experience.

The host says he has heard of AI companies engaging in what he calls “token dumping”—for example, selling $100 worth of model usage for $50.

In his framing, such businesses may effectively become money-losing token resellers.

He asks whether Lovable is profitable or close to profitable.

Anton does not give a direct profitability figure.

Instead, he says Lovable always monitors its margins but prioritizes what produces the best result for customers.

He specifically rejects the idea of automatically choosing a cheaper model when Lovable can measure that the cheaper model is meaningfully worse.

The company’s decision criterion is customer performance, not minimizing model cost at any price.

The host asks whether subscriptions are unlimited.

Anton says they have caps.

Customers who exhaust their included usage can top up, and Lovable offers multiple subscription arrangements.

Asked how many customers exceed their allowance, Anton says that on the lowest subscription tier, approximately 60% of customers do so.

The host interprets this as evidence that customers are willing to pay substantial overages because the economic value created by the product is so much greater than its price.

The host gives a hypothetical comparison.

Even if somebody eventually spent $6,000 per year because of heavy token usage, he argues that this could remain trivial compared with software that previously cost approximately $500,000 to develop.

His point is that customers should not focus narrowly on token spending if AI has reduced total development cost to somewhere around 0.1%–1% of the previous alternative.

Anton agrees with the broader principle.

He says the defining requirement for customers is speed and that using more AI usually allows them to move faster.

Therefore, in many cases, the additional AI spend is economically justified.

The host describes a new organizational behavior emerging because software has become so cheap to create.

Two different employees or groups independently created versions of an intranet:

• one for Japan;
• another for the United States.

They did not fork a common codebase.

They simply created two separate Lovable projects.

The host asks whether this apparent duplication is actually desirable.

Traditionally, a company would probably create one software project and attempt to reconcile all departmental requirements inside it.

With AI development, however, two capable people can make two independent attempts quickly and cheaply.

That effectively gives the company “two swings at bat.”

Anton says he is a strong supporter of very rapid experimentation.

He gives an example from time he spent at CERN in Central Europe.

There, he encountered a concept he calls “co-opetition.”

According to his description, two relatively isolated teams could work on aspects of the same particle accelerator without sharing their results until publication.

That structure allows each team to develop independently.

Afterward, the organization can compare outcomes and learn from the different approaches.

Anton argues that this reduces the risk that everyone converges prematurely on the same local minimum.

He compares the mechanism to:

• competition in free markets;
• parallel work in academia.

Because engineering effort itself is becoming less of a bottleneck, the scarce resource shifts toward determining what the right thing to build actually is.

Therefore, if an organization has enough capable people, Anton sees value in asking several of them to solve the same problem independently.

Anton says that after creating parallel Lovable projects, a user can ask the platform to inspect another project, identify several desirable elements and bring those features into the preferred version.

The user can then run a split test or experiment to determine whether the imported changes improve the relevant customer metrics.

This turns software development into a much more iterative selection process:

• create multiple approaches;
• compare them;
• combine the strongest elements;
• experimentally measure which version produces better outcomes.

The host says somebody used something transcribed as “Fable” to build Fortnite or a Fortnite-like project.

Anton says he has seen some of the 3D games.

The host asks about the latest Anthropic version, again transcribed as “Fable.”

Anton says Lovable uses it as one of the models powering the platform.

The precise model/product name is unclear and must be verified.

Asked whether the newer model represents a major step forward, Anton says it can create highly sophisticated and visually impressive work on the first attempt.

He specifically cites beautiful 3D games as an area where increased intelligence is obvious.

However, he argues that the core product-development problem does not disappear.

As a project evolves, humans still need to work with the agent to determine:

• what should be built;
• how the work should be planned;
• what strategic direction makes sense;
• what experiments should be run;
• what data should inform those decisions;
• which outcomes matter to the business.

Anton therefore distinguishes between execution intelligence and strategic judgment.

Model intelligence is improving rapidly at producing sophisticated artifacts.

The harder problem—deciding what is worth building and what direction the business should take—is improving less quickly.

Human beings who understand how to use the AI, provide the correct data and make good strategic choices therefore remain central.

The host says he likes both the product and Anton as a founder, but says the most important aspect is the business outcome produced by the technology.

He strongly recommends Lovable to listeners and says businesses should simply put it on a corporate card and begin building.

He jokes that Lovable has repeatedly been declared dead because new foundation models keep improving.

At the same time, he says Lovable repeatedly adds another roughly $100 million in revenue, studies its customers and continues surviving and growing.

He congratulates Anton for that entrepreneurial performance.

Anton thanks Jason and says he enjoyed the conversation.

The closing exchange references Paris and the Palace of Versailles, followed by a joke that one day something comparable might be built with Lovable and “optimist robots” [source wording; verify if needed].

Two non-interview promotional segments appear in the transcript.

The ad claims that if one were designing a new global financial system from first principles, it would not rely on decades-old payment rails. Airwallex is presented as an AI-native platform combining global accounts, cards and payments, intended to make international business operate more like a local market. The ad contrasts this with companies allegedly adding AI to legacy infrastructure.

The ad argues that important conversations often happen away from desks and promotes a wearable/hands-free recording device transcribed as “Plaud.” It claims the associated AI converts recordings into notes and next steps, allowing users to remain engaged in conversations rather than manually taking notes.

These are sponsor claims and should not be treated as independently verified editorial assertions.

Speaker: Pat
Exact quote: “Fundamentally, this is a technology business.”
Context: Pat argues that Intel lost its way when technical leadership was displaced by business-oriented leadership.
Topic: Technical leadership / Intel strategy.

Speaker: Pat
Exact quote: “When you're making these hard core technical, you know, decisions that affect billions of dollars, you don't do that through a spreadsheet.”
Context: Explanation of why technology trends can justify investments that look financially unattractive in isolation.
Topic: Technology investment / executive judgment.

Speaker: Pat, recounting Steve Jobs
Exact quote: “I've been working on that the last four releases.”
Context: Jobs reportedly told Intel that Apple had already been porting its operating system to x86 before the public transition.
Topic: Apple strategy / strategic preparation.

Speaker: Pat
Exact quote: “No, I can supply myself better.”
Context: Pat’s characterization of Apple’s reasoning for moving toward its own chips rather than simply viewing Intel as a failed supplier.
Topic: Apple Silicon / vertical integration.

Speaker: Pat
Exact quote: “These aren't just graphics cards anymore. These are general purpose computing devices that can start applying to these other workloads.”
Context: Pat describes the realization that GPUs could be used for HPC and eventually many non-graphics workloads.
Topic: Nvidia / GPU computing.

Speaker: Pat
Exact quote: “I don't care whose chip it is. I don't care what you're designing. I'll be your manufacturing partner.”
Context: Pat summarizes the strategic difference between TSMC’s foundry model and Intel’s proprietary IDM model.
Topic: TSMC / semiconductor manufacturing.

Speaker: Pat
Exact quote: “That should just put a chill in everybody's spine.”
Context: Reaction to his claim that Taiwan has less than three weeks of energy reserves.
Topic: Taiwan / semiconductor supply-chain risk.

Speaker: Pat
Exact quote: “Never do you need to do anything, a shot to be fired.”
Context: Pat argues that an energy blockade alone could disable Taiwan’s fabs and have enormous global economic consequences.
Topic: Taiwan blockade risk.

Speaker: Pat
Exact quote: “I'm an optimist, you know, that we're in a couple of decade buildout.”
Context: Pat argues that the economic value of additional AI intelligence could sustain demand far longer than a normal technology bubble.
Topic: AI infrastructure / long-term outlook.

Speaker: Pat
Exact quote: “I have to make AI 10,000 X better.”
Context: Pat says AI needs orders-of-magnitude reductions in token cost and energy use to unlock mass adoption.
Topic: AI efficiency.

Speaker: Pat
Exact quote: “There has not been a time in human history where it's been better to be a technologist than the one we're in right now.”
Context: Pat predicts major breakthroughs across chemistry, language, materials, medicine and poverty reduction.
Topic: Technology outlook.

Speaker: Pat
Exact quote: “Every time we have one of those corrections, say thank you.”
Context: Pat argues that periodic market corrections prevent AI valuations from becoming even more speculative.
Topic: AI valuations / market cycles.

Speaker: Pat
Exact quote: “This decade.”
Context: Direct answer when asked when quantum computing will begin producing meaningful results.
Topic: Quantum computing.

Speaker: Anton
Exact quote: “I talk about empowering humans.”
Context: Anton describes Lovable’s larger mission beyond simply generating software.
Topic: Lovable mission.

Speaker: Host/Jason
Exact quote: “You come to build the software, but you stay to build the business.”
Context: Summary of Lovable’s expansion from app creation into hosting, operations and AI business assistance.
Topic: AI co-founder / business platform.

Speaker: Anton
Exact quote: “It's the humans knowing how to use the tool to get and to plug in all the right data, to be able to take the right decisions.”
Context: Anton says improved models can generate sophisticated artifacts but do not eliminate strategic product and business judgment.
Topic: Human role in AI development.

Speaker: Anton
Exact quote: “I'm actually a huge fan of very rapid experimentation.”
Context: Anton endorses allowing several teams to build competing versions of the same software now that engineering is relatively cheap.
Topic: Product development / experimentation.

Speaker: Anton
Exact quote: “Since the engineering is less of the bottleneck, it's more the question of what is the right thing to build.”
Context: Anton argues AI shifts scarcity from implementation toward product judgment.
Topic: Future of software engineering.